An open resource from AI Swiss

BASE: shape AI around your work

BASE on video

1 min 46 s · in French with French subtitles · produced with InnovaudDemonstration using Claude Code. Other AI tools work too.

Read the transcript

This transcript is translated from the French original.

[Marie-Laure Gallez, from Innovaud, speaks to the camera, then a team of three people talks around a table.]

Companies are told to use artificial intelligence, but many still wonder what for, in concrete terms, and also how. It isn’t the will that’s lacking, it’s mainly the starting point. And that starting point is found in conversations about the ways we would like to work, about the team’s pain points, about repetitive tasks. And that is independent of AI.

These conversations can be recorded and given to an AI guided by BASE. BASE is a practical method developed by AI Swiss and supported by Innovaud to go further in how you use AI.

[Charles-Édouard Bardyn, co-founder of AI Swiss, faces the camera. Three words appear: converse, analyse, simplify.]

Today, in our tools, we tend to configure the behaviour we want with clicks and prompts. There is an alternative. You can simply describe your needs in your own words, with your own language and your knowledge of your trade. Let the AI handle the mechanics. And you end up with an assistant that meets your needs, one you have shaped entirely through language.

[On screen, the meeting transcript is handed to the AI, which replies with a brief: each task considered is marked yes, later or no, with its reason.]

From this transcript, the AI can generate proposals about what is feasible and what is not. Proposals that you are free to accept, reject or change. The result: guided by BASE, the AI builds your assistant in folders on your computer, which you own. Everything is readable, everything is transparent, and you can correct whatever you want through language.

[Three words appear on the computer screen: readable, transparent, modifiable.]

Now comes the moment of truth. Let’s launch the assistant to check whether it has followed the needs properly. In any case, everything remains modifiable.

And now it’s your turn! Visit www.a-i.swiss to make it your own and shape AI to your needs.

[The address www.a-i.swiss appears, followed by the Innovaud and BASE logos.]

What you teach an AI tool stays tied to that tool: in a conversation, in a product’s settings, in files organised its own way. Switching tools often means starting again. BASE is for everyone, individually or as a team, so that you can work with AI without losing control over what it knows about you and how it uses that knowledge.

You describe your work and what you expect from AI. Guided by BASE, your AI tool shapes itself around that description: it relies on your rules, consults your sources, prepares the checks you have defined and leaves the decisions to you. Your method becomes a set of readable files that you keep and that different AI tools can use.

Openly licensed and free of charge. No account or subscription. Access to an AI model is at your own cost.

The case in the video

The Invest team helps foreign companies set up in the Canton of Vaud. After its filmed discussion, the AI guided by BASE proposes tasks to the team. The team answers each one.

Prepare follow-ups after an event

A draft message for each contact met. The person who was there reviews and sends it.

Team’s answer:Yes

Keep the contacts file up to date

First, management must decide whether the assistant may access this file.

Team’s answer:Later

Spot companies that set up without the team knowing

A simple comparison tool against the commercial register will do better than AI.

Team’s answer:No

How it works in practice

Picture a property management company. Every morning, its team has to sort dozens of messages: a water leak, a heating breakdown, a lease termination or an administrative request. Only one person really knows the priorities, the exceptions and the right people to call. That person is retiring soon. The team wants to pass on this knowledge and get an assistant that prepares the sorting, without sending messages or making decisions on the team’s behalf.

  1. The discussion

    The team talks for an hour about its work, its difficulties and what must stay under its control. The conversation, recorded with the consent of those present, is transcribed, then handed to an AI tool with this request: “With BASE, show us what an assistant could do for our team.”

  2. The proposal

    The tool produces a brief, written in the team’s own terms, that summarises what it has understood, presents the possible tasks, and specifies what the assistant will never do and what AI will not solve. The team reviews each proposal and answers yes, later or no.

  3. The build

    The tool then creates the folder that matches the approved needs: the levels of urgency, the rules that apply to lease terminations, the tradespeople to contact for each building and the decisions reserved for management. It then puts the assistant to the test on cases mentioned during the discussion, to check that it responds the way the team expects.

  4. Daily work

    The next day, the team opens the assistant in its AI tool, drops the export of the morning’s messages into the folder and asks: “Sort the technical inbox.” Urgent items come first, messages about the same leak are grouped together, and replies are drafted in each tenant’s language. Someone reviews them and sends them: management has chosen not to connect the assistant to any messaging system.

  5. Adjustment

    As the days go by, the team keeps shaping the assistant through conversation. A rule is missing? The team says so: “For any leak, also notify the building’s caretaker.” Guided by BASE, the tool shows the proposed change and saves it only once it has been approved. From then on, the new rule applies to the requests it covers.

With BASE, you no longer have to work out which assistant, procedure or document fits each request. You state your intention: your assistant’s files, created with BASE, tell your tool how to proceed, what information it needs and which checks to run. You review the result.

How to start

With an AI tool that reads your folders (recommended)

  1. Get BASE onto your computer: download base.zip and unzip it. If you use Git, clone the repository instead to keep up with updates.
  2. Put your document in a folder of your own: the transcript of a discussion, a procedure or your notes.
  3. Start a new conversation in an AI tool that can read files on your computer, such as the Claude Code app or Codex, and tell it where the BASE folder is.
  4. Tell it:“Here is what we have. With BASE, help me structure it. Reply in English.”

As in the video, the tool first gives you a brief: for each task, you answer yes, later or no. Before your answers, only this brief is written. Then the tool shows you the files it plans to create and waits for your approval.

Only have a web chat, such as ChatGPT or Claude?

You can still see what BASE would make of your work, then discuss it with your AI to understand how it works. Building the assistant will then need a tool that opens a folder.

  1. Attach a document in your AI chat: the transcript of a discussion, a procedure or your notes.
  2. Paste this request:“Read https://raw.githubusercontent.com/ai-swiss/base/main/PARTIR-DE-CE-QUE-VOUS-AVEZ.md and follow it to show us, based on the attached document, what an assistant could do for us. Start with a proposal and do not build anything until we agree. Reply in English.”

If your chat cannot read this address, the guide Try BASE without installing BASE provides the file to attach with your documents. Only send it documents that you can entrust to an AI provider.

Try it on a fictional example

To see a BASE assistant at work first, try a fictional example whose answer can be checked. In the BASE folder, open exemples/assistant-devis-demo in an AI tool that can read the files on your computer, such as Claude Code or Codex.

This quote assistant requires two signed engagements before granting a loyalty discount, whereas the client record for Dupont SA mentions only one. Ask: “Is Dupont SA entitled to the loyalty discount?” The assistant must consult both files, answer no and point out that any exception must be approved by a person.

With a chat such as ChatGPT, without installing anything, the guide Try BASE without installing BASE offers a simpler trial with the same quote assistant: the chat receives the assistant as a single document and reads no files on your computer.

Why AI Swiss leads this project

AI Swiss is a non-profit association. In our work with AI, we see two ways of doing things. The first leaves it to a product to decide how information is found and used, without always showing how. In the second, that information stays visible, in files that the team can read, correct and take with it. BASE provides the means to follow this second approach.

BASE formalises practices that emerged from work in the field and were refined through use. It makes them reproducible and adaptable without depending on the association.

A word on sovereignty

A model developed in Switzerland, even an open one, does not make anyone sovereign on its own. It still has to be installed on an infrastructure, connected to your files by a tool and often integrated into a platform that may be proprietary. An AI solution rests on a chain of links: the model, the infrastructure that runs it, the tools and platforms that make it usable, and then what you know and know how to do. Being sovereign requires being able to choose each of these links and, if need be, replace it without having to rebuild the others.

BASE focuses on one link, your knowledge and know-how: the longest to build up, the most costly to rebuild and the only one that no provider can sell you, because it is your own. BASE guarantees nothing about the others: where the model runs, which data its provider can access and the terms that apply depend on the tool and the contract you choose. A Swiss or open model has its rightful place in that chain, as one link among others.

What you keep

Your knowledge and working methods are written down in text files, on your computer or on your organisation’s server. These files can set out:

Members of your team can open these files and understand how the work is organised. The AI relies on the same reference.

BASE is a starting point: a skeleton with the minimum structure an assistant needs to stay readable and verifiable, and one that grows with you.

You flesh it out as your needs grow, without changing method when one person becomes a team, then an organisation: an owner and a review date for each document, validation rules for each team, your own fields. The format is open, and you are free to take it and extend it. Overall security remains your organisation’s responsibility: BASE replaces neither its access controls nor its reviews.

Why give AI a map

Giving access to all your documents is not enough. Try it: create seven text files, hide a joke in one of them without pointing it out, then ask an AI tool to find it. To be certain of finding it, the tool will have to examine every file. Across an organisation’s thousands of documents, reading everything is no longer reasonable: each read uses computing power, takes time and consumes resources. So the tool reads some of the files, in a few seconds, and may claim to have understood the whole set. Without having read everything, it does not have the full picture, and it can miss the useful information without saying so.

  1. The aisle nameThe folder map
  2. The shelf labelThe domain
  3. The back-cover blurbThe header of each item
  4. The bookThe content, opened only if it is useful

A library solves this problem without anyone having to read every book: the aisle name, then the shelf label, then the back-cover blurb lead to the right book. BASE applies the same principle to your folder: the header of each item states what it covers, when it applies and when it does not. Your tool first reads these signposts, then opens only the items useful for the request. If no item covers the request, or if two remain possible, the map shows it and the tool is instructed to ask a question rather than guess.

The map does not write itself. BASE tells the tool how to build it and keep it up to date with you. You describe your work in your own words. The tool proposes a structure, which you correct before adopting it as the reference.

For people who already use instruction files or skills

Several AI tools already read an instruction file when a folder is opened, as well as skills, that is, procedures written for AI, each with a summary that helps the tool select it. BASE extends them without replacing them.

It declares the nature of each item: what guides the AI is kept separate from the information it should only consult, and know-how is kept separate from knowledge. The map helps the model choose the relevant item, or hold back when the request is not covered or remains ambiguous. A BASE router lets you test these decisions again after each change.

Safeguards, such as presenting a change before it is saved, are applied by code when the action goes through BASE, and the documentation states whether a guarantee rests on code or on an instruction. Finally, from a common source, BASE produces the entry file each supported tool expects, such as AGENTS.md, CLAUDE.md or Cursor rules. The whole team can thus work on the same folder, whatever tool each person uses.

Questions people ask us

Do you need to get all your data in order first?

No. You do not need to reorganise all of your organisation’s information before you start. Start from a specific task and from what you already have: a discussion, a procedure, a few document templates or your own notes.

BASE helps you structure what is useful for that task and make visible the information that is missing, contradictory or still to be confirmed. It does not turn missing or outdated information into reliable knowledge. You remain responsible for what is authoritative.

Is it too technical for an SME without IT staff?

Day to day, no: you write to the AI as you would to a colleague. Everything happens in a chat window with the AI you have shaped using BASE, in the tool of your choice. You ask for what you want to achieve, review the result and phrase your corrections in the language of your trade. There are no menus to learn and no workflows to configure. A turnkey product often requires you to adapt your work to its settings. With BASE, you describe the way you work.

Setting it up requires a tool that can read and modify files on your computer, such as Claude Code, Codex, Cursor, GitHub Copilot or OpenCode, as well as Node.js, a free program that the tool can help you install. Depending on the tool, this means opening an application or copying a command. A first trial can be done in an AI chat, without installing anything. Setting it up may require some initial help, and BASE is not a ready-to-use hosted service: choosing and configuring your AI tool is up to you.

Can you change models or use several tools?

Yes, provided those tools can read the files or connect to BASE. You can work with Codex and another member of your team with Cursor, while in-house software consults the same rules and the same sources.

Models are not interchangeable. BASE requires a model that can use tools and read and modify files. Many recent models, proprietary or open, clear that threshold. Even then, they may interpret an instruction differently and do not all have the same capabilities. A change therefore calls for adjustments and checks. The advantage is that you keep the reference method instead of having to rebuild it in a new interface.

What happens to confidential data?

The core of BASE does not transmit any data by itself. It works with the files placed in the storage location you have chosen. That does not mean everything stays where it is: your AI tool may send the content of a conversation or of a file it opens to its provider. Three BASE options, switched off until you enable them, also call on a model you choose (semantic routing, evaluation, Studio workshop chat).

You can mark a document as confidential, or declare an entire folder “local only”. When your tool goes through the BASE MCP server, or through a BASE option that calls a model, that content is not sent to a hosted model. A tool that reads the files directly is not stopped by this marking.

Before using personal, medical or financial data, or data subject to professional secrecy, check the tool you have chosen, its contractual terms, where processing takes place and your organisation’s rules. BASE does not replace access management or your compliance obligations.

How much does it cost, and who is behind it?

BASE is free of charge, published under open licences, with no account or subscription. To use it, you need access to a model, a lightweight AI tool that lets that model read and modify files, and Node.js, which is available free of charge. Access to the model, the tool and any help you may need with setting it up are at your own expense.

BASE is led by AI Swiss. Innovaud, the innovation promotion agency of the Canton of Vaud, is a partner in the project and helped to launch the trade-specific examples intended for SMEs. The association sells neither BASE subscriptions nor a hosted platform.

What BASE does not promise

A written instruction steers the model but does not constrain it. AI can make mistakes, ignore a rule or produce an answer that is convincing but wrong. Keep control over important decisions and adapt your checks to their possible consequences.

BASE’s code can apply certain safeguards, in particular requiring a change to be presented before it is saved and limiting access to the authorised folder. These safeguards only apply to actions that go through the corresponding mechanisms. If the tool acts directly on the files, they depend on the tool’s own settings.

BASE does not automatically make your organisation secure or compliant. It does not replace identity and access management, data leak prevention, legally required archiving or the controls required in regulated environments.

BASE combines a proposed open standard with a reference implementation. The standard describes how to organise an assistant’s files; the implementation helps to create them, check them and make them usable by different tools. The code is published under the Apache 2.0 licence, the documentation and examples under the CC BY 4.0 licence. Use cases in partnership with Innovaud.